Re: 2.6.9-rc1-mm4

[Nick Piggin <nickpiggin@yahoo.com.au>]

Martin J. Bligh wrote:
>>>Well, the good news is that it compiles now, and without forcing ACPI on.
>>> Yay!
>>
>>Does it boot?
> 
> 
> Yup. Performance is the same as other -mm's (scheduler changes bring it
> down from mainline quite a bit, but otherwise OK).
> 
> Kernbench: (make -j N vmlinux, where N = 16 x num_cpus)
>                               Elapsed      System        User         CPU
>                 2.6.9-rc1       44.97       98.66      576.77     1501.33
>             2.6.9-rc1-mm1       46.92      107.27      594.10     1493.67
>             2.6.9-rc1-mm2       46.95      107.80      593.65     1493.33
>             2.6.9-rc1-mm4       46.93      108.91      593.19     1495.00
> 

I'm looking into this performance thing a bit now (although I think
Andrew is going to drop nicksched from the next mm).

It doesn't seem to be from lack of timeslice: with the default
timeslice, nicksched gives average timeslices for a make -j vmlinux
roughly the same size as those for the 2.6 scheduler (22ms).

Increasing base_timeslice can get it up to more than 25% (28ms) larger,
but it still isn't as quick.

I suspect it may be unfairness in the 2.6 scheduler improving cache
utilisation. Need to find a way to measure that though :P